How Much MRI Preprocessing Is Enough? A Cost-Utility Study for Brain MRI Foundation Models

A systematic study of brain MRI foundation models finds that heavier preprocessing does not consistently improve performance, and that the lowest stable preprocessing level preserves most utility across several clinical tasks while requiring far less computation. The work, posted to arXiv on 6 June 2026, compared a graded P0–P7 preprocessing spectrum for two self-supervised learning methods—masked autoencoding (MAE) and joint-embedding predictive learning (JEPA)—trained on 20,000 heterogeneous brain MRI volumes [1]. The authors kept the 3D ViT backbone, masking protocol, and downstream evaluations fixed to isolate the effect of preprocessing intensity [1]. Levels P0 and P1 proved numerically unstable, making P2 the lowest-cost feasible level [1]. Beyond P2, choosing the best feasible preprocessing level improved aggregate utility by only 3.4 percentage points for MAE and 1.8 percentage points for JEPA, and most paired comparisons between P2 and higher levels were statistically unresolved [1][2]. The benefit of stronger preprocessing was task-dependent rather than monotonic. IDH mutation prediction showed modest gains, brain-age regression was often near or best at P2, and GLI/PED tumor segmentation was already maximized at P2 [1][2]. Mild cognitive impairment (MCI) classification was the clearest setting in which heavier preprocessing provided meaningful utility: MCI linear probing gained 9.9 percent and 32-shot classification gained 8.5 percent, both achieved by P7 models that required approximately 18.3 times the preprocessing cost of P2 [2][3]. The authors also demonstrated that much of the P7 advantage for MCI could be recovered by applying stronger preprocessing only during downstream fine-tuning, without requiring P7 throughout pretraining [1][2]. This cross-level transfer result suggests that the cost of heavy preprocessing can be deferred to task-specific stages rather than borne across the entire pretraining pipeline. These findings align with earlier work questioning the value of extensive MRI preprocessing. A 2022 study on brain tumor segmentation reported that the most popular standardization steps added no value to neural network performance and could even hamper it, identifying voxel-spacing unification as the only essential transformation [4]. A separate review of public brain MRI datasets found that standard preprocessing—including bias-field correction, intensity normalization, skull stripping, and spatial registration—increased internal consistency within datasets but did not fully remove inter-dataset differences; 83.89 percent of features still exhibited statistically significant shifts after Bonferroni correction [5]. The new study recasts MRI preprocessing as a downstream-aware cost–utility decision rather than a default escalation pipeline [1]. Code is available at the project’s GitHub repository [1].